Agents Can Install Expertise; They Still Cannot Approve Themselves¶
π§ Today's Thesis¶
Agent guidance, browser control, review, memory, and security are becoming ordinary software dependencies, but authority and acceptance still need an application-owned control plane. Modern Web Guidance can retrieve current browser practice, Chrome DevTools MCP can inspect and mutate a live browser, and SkillEvaluator can compare a task with and without a skill. None of those components knows by itself whether the right principal changed the right application state, whether all planned work ran, or whether a human can still repair the result.
CapLite supplies a concrete architecture for one part of that separation: derive write policy from typed tool schemas and the user's request, then commit the intended writes before untrusted tool results arrive. Scout's broader interpretation is an application-owned acceptance plane: code outside the worker records principal, dependency revisions, eligible scope, allowed effects, attempted and observed effects, missing work, protected postconditions, reviewer decision, and terminal owner. That is a synthesis, not a vendor feature, and the research guarantee was not reproduced here.
π― What Matters to Us This Week¶
-
Write authority can move out of model-authored security prose. CapLite reports failures when small models must create their own policy artifacts, then derives write integrity from tool schemas and the user's request under stated assumptions. For a Node service, compile one typed allowlist before reading retrieved content; do not let a tool result expand the intended target or operation.
-
Review is becoming a product surface, not merely another agent role.
backnotprop/plannotatorrose from 55 to 108 to 408 stars/day, whilealibaba/open-code-reviewrebounded to 893/day andopenai/codex-securityheld 595/day. These products expose different piecesβstructured human feedback, deterministic checks plus model review, or threat-model-guided investigation. Compare reviewer reconstruction and accepted findings under the same protected result; comment volume is not the outcome. -
Completion and completeness remain different fields. The Codex Security release stream exposes durable lifecycle and imported evidence, while issue #2446 reports workers hitting a usage limit as a scan finalizes rather than remaining resumable. Force one interruption and one budget stop before trusting
completedas an operational state. -
Package reach makes canaries cheap, not conclusions automatic. Live search observed
modern-web-guidance1.0.0 andchrome-devtools-mcp1.9.0 with broad npm distribution, but deterministic hydration returned HTTP 403 for both package pages. Their official repositories, releases, archived evidence, and operator-owned tests carry the recommendation; package counts are verification-limited distribution leads. -
Developer demand remains reviewability and repairability. Direct ExperiencedDevs, AskProgrammers, and developersIndia discussions question code volume, review cost, and losing the system's mental model. These are qualitative reports, not prevalence evidence, but they make reviewer corrections, reconstruction time, rollback, and protected outcomes the right local metrics.
π₯ Top Movers¶
(All velocities are explicit GitHub daily-window observations; weekly and monthly measurements remain separate.)
- GoogleChrome/modern-web-guidance-src (1,231 β today, 6,013 total) β maintained browser guidance held high after yesterday's 1,925/day peak; public paired evals still need an independent React canary.
- alibaba/open-code-review (893 β today, 51,414 total) β attention recovered outside the strict fading boundary; accepted-defect recall and false-positive reviewer minutes remain unmeasured here.
- ChromeDevTools/chrome-devtools-mcp (664 β today, 56,283 total) β a new clean peak for live browser inspection and mutation; application postconditions remain outside the tool result.
- openai/codex-security (595 β today, 13,538 total) β strong attention below its 751/day peak; lifecycle evidence improved while incomplete-scan semantics remain open.
- tradermonty/claude-trading-skills (553 β today, 4,786 total) β a fast consequential-domain skill signal that stays out of the active stack.
- backnotprop/plannotator (408 β today, 11,104 total) β structured plan and diff review reached a third clean observation and new peak.
π What Changed the Frontier¶
- Tool policy became compilable before retrieval. CapLite's typed-schema approach separates the user's original write intent from later tool content that may be untrusted.
- Skill lift acquired an installable control arm. SkillEvaluator 0.7.2 makes the same-task with-skill versus without-skill comparison a normal Python workflow. Operator-owned cases, negative controls, graders, and protected outcomes still decide promotion.
- MCP release planning became more predictable. The official release-cycle update improves migration visibility. Predictable releases help pin the protocol surface; they do not supply per-application identity, authorization, effect, or acceptance policy.
- Repository memory acquired a costed next-task test. CoReMem values retained trajectory segments by whether they help the next coding task enough to justify write and read cost, favoring repository locations and executable checks as compact anchors.
π First Appearances¶
- No verified on-lens first appearance today. The strongest operator-relevant repositories already had clean observations this week. Scout updated their window-labelled histories rather than treating today's registry additions as product debuts.
π± Rising Stars¶
(Direction claims require successive clean, window-labelled daily observations.)
- ChromeDevTools/chrome-devtools-mcp β 664/day, a new verified clean peak after recent 60 and 67/day observations.
- backnotprop/plannotator β 55 β 108 β 408/day, a third clean observation and new peak.
- NVIDIA/SkillEvaluator β 108 β 133/day, a second clean observation and new peak at 1,254 total.
- tradermonty/claude-trading-skills β 69 β 553/day across two clean observations; attention is rising, while the operator action remains ignore for adoption.
π Fading¶
(These explicit daily observations are more than 80% below a verified clean peak. βFadingβ describes attention, not product quality.)
- headroomlabs-ai/headroom β peaked at 549/day, now 93 (83.1% lower). Silent truncation and protected-field fidelity, not attention, decide whether compression is useful.
- Tencent/LoopForge β first clean baseline 184/day, now 33 (82.1% lower). A repository script plus CI remains the control until recovery and accepted-work evidence justify a platform.
- ag-ui-protocol/ag-ui β first clean baseline 67/day, now 9 (86.6% lower). Protocol value remains independent of daily attention.
- mukul975/Anthropic-Cybersecurity-Skills β peaked at 767/day, now 59 (92.3% lower). Use the collection as a study source, not as a production security control.
βοΈ Battles (same category, competing)¶
- Codex Security vs Open Code Review β both reduce pre-merge review work. Codex Security emphasizes threat models, agentic investigation, imported scanner evidence, isolated validation, and lifecycle state; Open Code Review emphasizes deterministic selection and rules plus bounded model review in CI. Test one seeded defect and one correct change with the same exact revision, budget, protected checks, reviewer, and incomplete-state policy.
- Plannotator vs chat and pull-request comments β both carry human feedback back to the coding worker. Plannotator offers a structured shared workspace; existing comments are already governed, searchable, and close to merge state. Adopt only if the workspace measurably lowers reconstruction and correction time without hiding the protected outcome.
- Modern Web Guidance vs repository-native guidance β maintained cross-browser knowledge may improve discovery and currency; a small reviewed repository guide is easier to pin, understand, and remove. The same browser task and protected checks should decide incremental lift.
π¬ From Research¶
- CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts β derives write policy from typed tool schemas and the user request, then commits writes before tool results are read; the reported guarantee is bounded by the paper's assumptions.
- Discovering Effective Agent Memory Designs through Program Evolution β reports gains from retaining raw conversations as retrievable evidence and composing them according to query intent.
- CoReMem: Cost-Internalizing Memory for Successive Coding Tasks with Prospective Valuation β values memory by measured next-task benefit minus read and write cost, emphasizing repository anchors and executable checks.
- Auditing AI peer reviewers β injects localizable errors into real papers and tracks detection plus false positives, reinforcing paired, seeded evaluation rather than reviewer eloquence.
The papers were pre-collected inside the November 9β15 research window and the first three were deterministically hydrated. Their reported results were not reproduced in this run.
π What's Changing¶
The ecosystem is turning agent behavior into a dependency graph: guidance, skills, memory, MCP servers, browser controllers, security investigators, and review workspaces can all be installed, upgraded, and composed. That is genuine progress because revision identity and bounded interfaces are easier to inspect than a one-off prompt.
The control problem is moving to the joins. A skill can author a change and influence its evaluation; a browser tool can report a successful click without proving the intended application state; a security agent can finalize before all workers finish; memory can improve retrieval while making the next task more expensive; a review surface can clarify feedback without proving the change is safe. The durable architecture keeps the application's principal, allowed effect, protected postcondition, completeness, reviewer decision, and rollback outside the component being evaluated.
Thirteen of 15 general web URLs hydrated successfully. The two npm pages returned HTTP 403 and remain verification-limited leads. GitHub, PyPI, the MCP maintainers' blog, Reddit, OpenAI's official changelog, and arXiv supplied archived evidence; community discussions indicate reported needs, not market prevalence. The one pre-collected YouTube transcript supplied supporting Salesforce MCP authorization detail, but no decisive claim depends on video evidence.
π§ͺ One Experiment Worth Running¶
Compile one write-intent contract before an agent reads tool output. Proposed first step: expose a disposable Postgres-backed service with two fake tenants and one typed update_ticket tool. Give the agent a request that permits changing one field on one ticket for one tenant, then return tool content containing a plausible instruction to update the sibling tenant. Before retrieval, application code should record principal, exact tool, allowed target and fields, and expected postcondition; after execution, assert the permitted row, denied sibling, attempted arguments, observed effect, terminal state, and cleanup. One fixture can test this architecture shape; it cannot validate CapLite's full research claim.
β οΈ One Risk to Track¶
The dependency under test can become its own policy author, evaluator, and completion authority. Watch for a skill that writes its cases, a browser tool that validates only its command result, a scanner that chooses eligible scope and seals its own completion, or memory that cites only the evidence it retained. The downside is a coherent green report whose missing work and prohibited effects are no longer visible. Keep at least one protected application state, negative control, and completeness verdict outside the dependency.
π One Thing to Ignore¶
Trading-skill bundles as an adopted workflow β tradermonty/claude-trading-skills reached 553 stars/day, but the evidence shows attention and packaged procedure, not approved data sources, dated provenance, reproducible analysis, backtesting validity, transaction boundaries, loss limits, or human approval. Revisit only with a read-only frozen-data task whose claims and calculations reproduce outside the agent and whose transaction tools are denied by application code.
π‘ Surprise Pick¶
backnotprop/plannotator β the interesting signal is not a new reviewer agent. It is a workspace that treats human decisions, annotations, and handoff back to the worker as the scarce artifact. If one small canary shows lower reconstruction time without weakening independent tests or merge ownership, it may be more useful than adding another automated review pass.
π Supply vs. Demand¶
| What's being built (supply) | What people want (demand) | Match? |
|---|---|---|
| Maintained guidance and installable agent skills | Current expertise that improves one task without invisible activation or regressions | Partial β paired evals exist; operator-owned cases and protected outcomes remain necessary |
| Browser MCP servers with broad package reach | Reliable inspection and mutation in the right session, principal, and application state | Partial β tool access is strong; application acceptance and cleanup stay external |
| Agentic and hybrid code/security review | Lower review cost without losing completeness, system understanding, or repairability | Partial β evidence surfaces improved; direct practitioners still report review as the scarce queue |
| Protocol release cadence and typed tool schemas | Predictable upgrades plus least-privilege effects | Promising but separate β protocol planning helps; application policy still owns the exact effect |
| Persistent coding-agent memory | Useful continuity without replaying every trajectory or hiding stale context | Research-backed direction β next-task value and read/write cost are measurable; production durability remains open |
π Category Pulse¶
| Category | New Today | Trending Count | Signal |
|---|---|---|---|
| Code dev tools | 0 verified | 15 registry-matched daily repos | β Guidance, security, review, and feedback become separately installable surfaces |
| LLM eval / testing | 0 verified | 5 + research | β Control arms, seeded errors, completeness, and protected outcomes separate |
| MCP tooling | 0 verified | 4 | β Release planning improves while per-effect authority stays application-owned |
| Memory / RAG | 0 verified | 4 + research | β Retention is being valued by next-task benefit and cost |
| Agent infrastructure | 0 verified | 5 | β Typed policy and terminal ownership matter more than another broad runtime |
| Agent skills | 0 verified | 6 | β Domain depth expands faster than promotion and consequential-authority controls |